aquantia: Fix Tx queue hangups
authorIgor Russkikh <igor.russkikh@aquantia.com>
Mon, 25 Sep 2017 07:48:48 +0000 (10:48 +0300)
committerDavid S. Miller <davem@davemloft.net>
Tue, 26 Sep 2017 20:44:31 +0000 (13:44 -0700)
Driver did a poor job in managing its Tx queues: Sometimes it could stop
tx queues due to link down condition in aq_nic_xmit - but never waked up
them. That led to Tx path total suspend.
This patch fixes this and improves generic queue management:
- introduces queue restart counter
- uses generic netif_ interface to disable and enable tx path
- refactors link up/down condition and introduces dmesg log event when
  link changes.
- introduces new constant for minimum descriptors count required for queue
  wakeup

Signed-off-by: Pavel Belous <Pavel.Belous@aquantia.com>
Signed-off-by: Igor Russkikh <igor.russkikh@aquantia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/aquantia/atlantic/aq_cfg.h
drivers/net/ethernet/aquantia/atlantic/aq_nic.c
drivers/net/ethernet/aquantia/atlantic/aq_nic.h
drivers/net/ethernet/aquantia/atlantic/aq_ring.c
drivers/net/ethernet/aquantia/atlantic/aq_ring.h
drivers/net/ethernet/aquantia/atlantic/aq_vec.c

index 214986436ece5a103e6452218cf34719d223b820..0fdaaa643073afabda6d66c9ebe0ef9e3b7695ae 100644 (file)
 
 #define AQ_CFG_SKB_FRAGS_MAX   32U
 
+/* Number of descriptors available in one ring to resume this ring queue
+ */
+#define AQ_CFG_RESTART_DESC_THRES   (AQ_CFG_SKB_FRAGS_MAX * 2)
+
 #define AQ_CFG_NAPI_WEIGHT     64U
 
 #define AQ_CFG_MULTICAST_ADDRESS_MAX     32U
index bf26a59a9d8e2ea39aac942985ac40353ee7a1fd..072a55029f047ba3652434e9d1e34b28c1df7997 100644 (file)
@@ -119,6 +119,35 @@ int aq_nic_cfg_start(struct aq_nic_s *self)
        return 0;
 }
 
+static int aq_nic_update_link_status(struct aq_nic_s *self)
+{
+       int err = self->aq_hw_ops.hw_get_link_status(self->aq_hw);
+
+       if (err)
+               return err;
+
+       if (self->link_status.mbps != self->aq_hw->aq_link_status.mbps)
+               pr_info("%s: link change old %d new %d\n",
+                       AQ_CFG_DRV_NAME, self->link_status.mbps,
+                       self->aq_hw->aq_link_status.mbps);
+
+       self->link_status = self->aq_hw->aq_link_status;
+       if (!netif_carrier_ok(self->ndev) && self->link_status.mbps) {
+               aq_utils_obj_set(&self->header.flags,
+                                AQ_NIC_FLAG_STARTED);
+               aq_utils_obj_clear(&self->header.flags,
+                                  AQ_NIC_LINK_DOWN);
+               netif_carrier_on(self->ndev);
+               netif_tx_wake_all_queues(self->ndev);
+       }
+       if (netif_carrier_ok(self->ndev) && !self->link_status.mbps) {
+               netif_carrier_off(self->ndev);
+               netif_tx_disable(self->ndev);
+               aq_utils_obj_set(&self->header.flags, AQ_NIC_LINK_DOWN);
+       }
+       return 0;
+}
+
 static void aq_nic_service_timer_cb(unsigned long param)
 {
        struct aq_nic_s *self = (struct aq_nic_s *)param;
@@ -131,26 +160,13 @@ static void aq_nic_service_timer_cb(unsigned long param)
        if (aq_utils_obj_test(&self->header.flags, AQ_NIC_FLAGS_IS_NOT_READY))
                goto err_exit;
 
-       err = self->aq_hw_ops.hw_get_link_status(self->aq_hw);
-       if (err < 0)
+       err = aq_nic_update_link_status(self);
+       if (err)
                goto err_exit;
 
-       self->link_status = self->aq_hw->aq_link_status;
-
        self->aq_hw_ops.hw_interrupt_moderation_set(self->aq_hw,
                    self->aq_nic_cfg.is_interrupt_moderation);
 
-       if (self->link_status.mbps) {
-               aq_utils_obj_set(&self->header.flags,
-                                AQ_NIC_FLAG_STARTED);
-               aq_utils_obj_clear(&self->header.flags,
-                                  AQ_NIC_LINK_DOWN);
-               netif_carrier_on(self->ndev);
-       } else {
-               netif_carrier_off(self->ndev);
-               aq_utils_obj_set(&self->header.flags, AQ_NIC_LINK_DOWN);
-       }
-
        memset(&stats_rx, 0U, sizeof(struct aq_ring_stats_rx_s));
        memset(&stats_tx, 0U, sizeof(struct aq_ring_stats_tx_s));
        for (i = AQ_DIMOF(self->aq_vec); i--;) {
@@ -240,7 +256,6 @@ err_exit:
 int aq_nic_ndev_register(struct aq_nic_s *self)
 {
        int err = 0;
-       unsigned int i = 0U;
 
        if (!self->ndev) {
                err = -EINVAL;
@@ -262,8 +277,7 @@ int aq_nic_ndev_register(struct aq_nic_s *self)
 
        netif_carrier_off(self->ndev);
 
-       for (i = AQ_CFG_VECS_MAX; i--;)
-               aq_nic_ndev_queue_stop(self, i);
+       netif_tx_disable(self->ndev);
 
        err = register_netdev(self->ndev);
        if (err < 0)
@@ -318,12 +332,8 @@ struct aq_nic_s *aq_nic_alloc_hot(struct net_device *ndev)
                err = -EINVAL;
                goto err_exit;
        }
-       if (netif_running(ndev)) {
-               unsigned int i;
-
-               for (i = AQ_CFG_VECS_MAX; i--;)
-                       netif_stop_subqueue(ndev, i);
-       }
+       if (netif_running(ndev))
+               netif_tx_disable(ndev);
 
        for (self->aq_vecs = 0; self->aq_vecs < self->aq_nic_cfg.vecs;
                self->aq_vecs++) {
@@ -383,16 +393,6 @@ err_exit:
        return err;
 }
 
-void aq_nic_ndev_queue_start(struct aq_nic_s *self, unsigned int idx)
-{
-       netif_start_subqueue(self->ndev, idx);
-}
-
-void aq_nic_ndev_queue_stop(struct aq_nic_s *self, unsigned int idx)
-{
-       netif_stop_subqueue(self->ndev, idx);
-}
-
 int aq_nic_start(struct aq_nic_s *self)
 {
        struct aq_vec_s *aq_vec = NULL;
@@ -451,10 +451,6 @@ int aq_nic_start(struct aq_nic_s *self)
                        goto err_exit;
        }
 
-       for (i = 0U, aq_vec = self->aq_vec[0];
-               self->aq_vecs > i; ++i, aq_vec = self->aq_vec[i])
-               aq_nic_ndev_queue_start(self, i);
-
        err = netif_set_real_num_tx_queues(self->ndev, self->aq_vecs);
        if (err < 0)
                goto err_exit;
@@ -463,6 +459,8 @@ int aq_nic_start(struct aq_nic_s *self)
        if (err < 0)
                goto err_exit;
 
+       netif_tx_start_all_queues(self->ndev);
+
 err_exit:
        return err;
 }
@@ -602,7 +600,6 @@ int aq_nic_xmit(struct aq_nic_s *self, struct sk_buff *skb)
        unsigned int vec = skb->queue_mapping % self->aq_nic_cfg.vecs;
        unsigned int tc = 0U;
        int err = NETDEV_TX_OK;
-       bool is_nic_in_bad_state;
 
        frags = skb_shinfo(skb)->nr_frags + 1;
 
@@ -613,13 +610,10 @@ int aq_nic_xmit(struct aq_nic_s *self, struct sk_buff *skb)
                goto err_exit;
        }
 
-       is_nic_in_bad_state = aq_utils_obj_test(&self->header.flags,
-                                               AQ_NIC_FLAGS_IS_NOT_TX_READY) ||
-                                               (aq_ring_avail_dx(ring) <
-                                               AQ_CFG_SKB_FRAGS_MAX);
+       aq_ring_update_queue_state(ring);
 
-       if (is_nic_in_bad_state) {
-               aq_nic_ndev_queue_stop(self, ring->idx);
+       /* Above status update may stop the queue. Check this. */
+       if (__netif_subqueue_stopped(self->ndev, ring->idx)) {
                err = NETDEV_TX_BUSY;
                goto err_exit;
        }
@@ -631,9 +625,6 @@ int aq_nic_xmit(struct aq_nic_s *self, struct sk_buff *skb)
                                                      ring,
                                                      frags);
                if (err >= 0) {
-                       if (aq_ring_avail_dx(ring) < AQ_CFG_SKB_FRAGS_MAX + 1)
-                               aq_nic_ndev_queue_stop(self, ring->idx);
-
                        ++ring->stats.tx.packets;
                        ring->stats.tx.bytes += skb->len;
                }
@@ -898,9 +889,7 @@ int aq_nic_stop(struct aq_nic_s *self)
        struct aq_vec_s *aq_vec = NULL;
        unsigned int i = 0U;
 
-       for (i = 0U, aq_vec = self->aq_vec[0];
-               self->aq_vecs > i; ++i, aq_vec = self->aq_vec[i])
-               aq_nic_ndev_queue_stop(self, i);
+       netif_tx_disable(self->ndev);
 
        del_timer_sync(&self->service_timer);
 
index 7fc2a5ecb2b7a01f594388e929c7affcb299af6f..0ddd556ff901c25682739a059673fa3a8398aed3 100644 (file)
@@ -83,8 +83,6 @@ struct net_device *aq_nic_get_ndev(struct aq_nic_s *self);
 int aq_nic_init(struct aq_nic_s *self);
 int aq_nic_cfg_start(struct aq_nic_s *self);
 int aq_nic_ndev_register(struct aq_nic_s *self);
-void aq_nic_ndev_queue_start(struct aq_nic_s *self, unsigned int idx);
-void aq_nic_ndev_queue_stop(struct aq_nic_s *self, unsigned int idx);
 void aq_nic_ndev_free(struct aq_nic_s *self);
 int aq_nic_start(struct aq_nic_s *self);
 int aq_nic_xmit(struct aq_nic_s *self, struct sk_buff *skb);
index 4eee1996a8259e561c15a17b9a3792ef79f280e2..02f79b0640baeabea52d43a2204363abcd3b6049 100644 (file)
@@ -104,6 +104,32 @@ int aq_ring_init(struct aq_ring_s *self)
        return 0;
 }
 
+void aq_ring_update_queue_state(struct aq_ring_s *ring)
+{
+       if (aq_ring_avail_dx(ring) <= AQ_CFG_SKB_FRAGS_MAX)
+               aq_ring_queue_stop(ring);
+       else if (aq_ring_avail_dx(ring) > AQ_CFG_RESTART_DESC_THRES)
+               aq_ring_queue_wake(ring);
+}
+
+void aq_ring_queue_wake(struct aq_ring_s *ring)
+{
+       struct net_device *ndev = aq_nic_get_ndev(ring->aq_nic);
+
+       if (__netif_subqueue_stopped(ndev, ring->idx)) {
+               netif_wake_subqueue(ndev, ring->idx);
+               ring->stats.tx.queue_restarts++;
+       }
+}
+
+void aq_ring_queue_stop(struct aq_ring_s *ring)
+{
+       struct net_device *ndev = aq_nic_get_ndev(ring->aq_nic);
+
+       if (!__netif_subqueue_stopped(ndev, ring->idx))
+               netif_stop_subqueue(ndev, ring->idx);
+}
+
 void aq_ring_tx_clean(struct aq_ring_s *self)
 {
        struct device *dev = aq_nic_get_dev(self->aq_nic);
index 782176c5f4f800f090353ee3184a98918dd1d0a3..24523b5ac68c1363f671ef0ac90fcd34ecf7bb95 100644 (file)
@@ -94,6 +94,7 @@ struct aq_ring_stats_tx_s {
        u64 errors;
        u64 packets;
        u64 bytes;
+       u64 queue_restarts;
 };
 
 union aq_ring_stats_s {
@@ -147,6 +148,9 @@ struct aq_ring_s *aq_ring_rx_alloc(struct aq_ring_s *self,
 int aq_ring_init(struct aq_ring_s *self);
 void aq_ring_rx_deinit(struct aq_ring_s *self);
 void aq_ring_free(struct aq_ring_s *self);
+void aq_ring_update_queue_state(struct aq_ring_s *ring);
+void aq_ring_queue_wake(struct aq_ring_s *ring);
+void aq_ring_queue_stop(struct aq_ring_s *ring);
 void aq_ring_tx_clean(struct aq_ring_s *self);
 int aq_ring_rx_clean(struct aq_ring_s *self,
                     struct napi_struct *napi,
index ebf588004c4677140934b152eeab13d4d3aecbf7..305ff8ffac2c9acc9633aa4fadde4a33d0392f8d 100644 (file)
@@ -59,12 +59,7 @@ static int aq_vec_poll(struct napi_struct *napi, int budget)
                        if (ring[AQ_VEC_TX_ID].sw_head !=
                            ring[AQ_VEC_TX_ID].hw_head) {
                                aq_ring_tx_clean(&ring[AQ_VEC_TX_ID]);
-
-                               if (aq_ring_avail_dx(&ring[AQ_VEC_TX_ID]) >
-                                   AQ_CFG_SKB_FRAGS_MAX) {
-                                       aq_nic_ndev_queue_start(self->aq_nic,
-                                               ring[AQ_VEC_TX_ID].idx);
-                               }
+                               aq_ring_update_queue_state(&ring[AQ_VEC_TX_ID]);
                                was_tx_cleaned = true;
                        }
 
@@ -364,6 +359,7 @@ void aq_vec_add_stats(struct aq_vec_s *self,
                stats_tx->packets += tx->packets;
                stats_tx->bytes += tx->bytes;
                stats_tx->errors += tx->errors;
+               stats_tx->queue_restarts += tx->queue_restarts;
        }
 }